What is it? The term “Hybrid Cloud” is generally used to describe an on-premise private cloud that operates in conjunction with public cloud. That’s NOT what Sensory is considering. Instead, we envision a “hybrid” of embedded or edge-based functionality with a cloud that can run on-premise, in your network, or in a public cloud.

Sensory’s Private Hybrid Cloud excels at running AI technologies, including wake words, voice commands, natural language understanding, biometrics, sound identification, computer vision and MUCH more. Most importantly, we are focused on maintaining privacy!

Why do it? With so many devices containing cameras and microphones, there is a growing privacy concern of personally identifiable information (PII) leaving the device.

The effects of this type of privacy loss are most commonly seen when you search for something and then immediately start seeing ads similar to that search. Another far more serious situation occurs when private data is leaked, hacked or phished, resulting in important financial or other private information ending up in the wrong hands.

Sensory’s on-device, embedded technology eliminates the risk of data theft by employing strictly edge-base processing. Processing data in this manner has the clear advantage of privacy, but it also has its downsides. Specifically, updating models is more difficult, and the memory and computing power required for state-of-the-art AI burdens products with unwanted costs. However, it’s possible to create a flexible hybrid architecture that allows the combined advantages of both client and cloud computing without risking privacy.

Architectures. Of course, Sensory can run edge-only or cloud-only solutions as these have been the standard approaches for AI companies. However, let’s look at some of the other architectures and how they add value:

Architecture 1: Feature extraction on-device and sending the features to the cloud for inference.

By performing feature extraction on-device and sending the features to the cloud for inference, the system achieves privacy, low cost of goods, ultra-low power consumption, and continuing improvements. This architecture is useful for wearables and cost or power-sensitive devices that need to be small.

Many DSP companies are producing AI focused inference engines that can reside on-device. These on-device decisions can still benefit in accuracy by way of revalidation in the cloud with a bigger and better engine, while still offering a high level of responsiveness.

Architecture 2: On-device processing of Inference & actions with feature extraction followed by decision revalidation in the cloud.

Moving the features into the cloud for revalidation, can also allow new models to be built, compressed, and sent back to the device. The on-device processing and power consumption costs do go up a little in these approaches, but there are huge advantages in model quality, and privacy, as well as always being available and responsive.

This sort of approach offers solutions for industries like automotive, where on-device intelligence and responsiveness without connection is essential, yet performance must also be state-of-the-art.

Sensory wants your feedback, we’re developing a modern approach with:

• Lightweight, dockerized app deployable on Amazon’s ECS, Azure’s Container Instances, Google’s GKE, or any Kubernetes hardware.
• Written in Go for high performance and versatility.
• Nvidia Triton backend to support deep learning frameworks such as Pytorch, Tensorflow, TensorRT, ONNX, and even custom backends.
• gRPC to facilitate fast device-to-cloud communication with very low latency.

Rather than going into “stealth mode” to work on the next big thing, we want to start having conversations now. Contact Us to discuss how Sensory’s cloud strategy can work for you.